Continual Refinable Network for Edge Device Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural network models face high memory/storage and computational costs, leading to low generalization and catastrophic forgetting when trained for sequential tasks, making them unsuitable for edge devices like smartphones or IoT devices, and causing degradation in network performance due to interference between old and new task learning parameters.
Innovation Solution
A continual refinable network (CRN) approach that directs gradients to flat local minima and uses a dynamic learning rate to refine network weights, maintaining a stability-plasticity trade-off and preventing catastrophic forgetting by dynamically updating shared learning parameters, thereby allowing the network to retain knowledge across tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a specific neural network is trained for every task, then task-specific performance is improved, but memory/storage and computational costs increase
Solution Approach 1:
The patent merges multiple task-specific neural networks into a single shared parameter space that can handle multiple tasks. Instead of maintaining separate networks for each task, the system uses one network with dynamically adjusted parameters, reducing memory/storage requirements while maintaining task-specific performance through parameter refinement.
Solution Approach 2:
The patent creates a universal neural network that can perform multiple different tasks by dynamically adjusting its parameters. The shared parameter space is designed to be multi-functional, allowing the same network structure to adapt to various tasks through parameter updates rather than requiring separate specialized networks.
2Quantity of substance
If a shared parameters space is used, then memory/storage costs are reduced, but catastrophic forgetting occurs
Solution Approach 1:
The patent implements dynamic parameter adjustment where the shared parameters are continuously refined based on the current task being performed. The system dynamically updates parameters to accommodate new tasks while preserving important information from previous tasks through controlled parameter evolution rather than static parameter sharing.
Solution Approach 2:
The patent changes the parameters of the shared neural network dynamically based on the task at hand. Instead of using fixed shared parameters, the system adjusts parameters adaptively to prevent catastrophic forgetting, allowing the network to maintain performance across multiple tasks by carefully managing parameter updates.
3Productivity
If new tasks are learned in a shared parameters space, then computational efficiency is improved, but interference between old and new task parameters causes catastrophic forgetting
Solution Approach 1:
The patent applies local quality by allowing different regions or aspects of the parameter space to have different properties. Some parameters are kept stable to preserve previous task knowledge, while others are allowed to change freely to adapt to new tasks. This selective parameter management reduces interference between tasks while maintaining computational efficiency.
Data Source
AI summary
A method includes accessing, using at least one processor of an electronic device, a machine learning model. The machine learning model is trained by directing a gradient direction of gradients to one or more flat local minima and using a dynamic learning rate for one or more additional tasks. The method also includes receiving, using the at least one processor, an input from an input source. The method further includes providing, using the at least one processor, the input to the machine learning model. The method also includes receiving, using the at least one processor, an output from the machine learning model. In addition, the method includes instructing, using the at least one processor, at least one action based on the output from the machine learning model.